A Reliability-Weighted Framework for Task-Aware Multimodal Perception in Simulated Coal Mine Roadways

Underground perception requires sensing modalities with different responses to dust, darkness, and interference. This study formulates a task-aware framework combining light detection and ranging (LiDAR), millimeter-wave radar, thermal imaging, and vibration sensing. The revised framework explicitly separates geometric correspondence, task-specific modality weights, feature-channel gating, spatial gating, and residual fusion. Mutual information is treated as a heuristic relevance prior, and posterior pooling is restricted to compatible classification outputs. Existing aggregate summaries report a detection score of 92.7%, equipment-state accuracy of 91.2%, and recognition accuracies of 82.9% and 90.2% at the highest separate dust and interference settings. These results support the feasibility of the complete framework in the reported controlled setting, while the revision avoids attributing the gains to a single adaptive component. LiDAR preprocessing takes 62.8 ms, exceeding the 50 ms system target. The manuscript therefore limits deployment claims and identifies matched ablations, independent validation, and complete performance profiling as priorities for further study.